feat: add router-level Prometheus metrics and centralize request tracking (#6146)

Signed-off-by: PeaBrane <yanrpei@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

lib/bindings/python/rust/llm/kv.rs

@@ -947,6 +947,7 @@ impl KvPushRouter {
             tokio::spawn(async move {
                 let mut stream = stream;
                 let mut first_item = true;
+                let mut first_token_gauges_observed = false;
 
                 while let Some(mut response) = stream.next().await {
                     // Inject worker_id into first response if tracker is available
@@ -957,6 +958,21 @@ impl KvPushRouter {
                         }
                     }
 
+                    // Observe per-worker TTFT/ISL gauges on first response with actual tokens
+                    if !first_token_gauges_observed {
+                        let has_tokens = response
+                            .data
+                            .as_ref()
+                            .map(|d| !d.token_ids.is_empty())
+                            .unwrap_or(false);
+                        if has_tokens {
+                            if let Some(ref tracker) = tracker {
+                                tracker.observe_first_token_gauges();
+                            }
+                            first_token_gauges_observed = true;
+                        }
+                    }
+
                     // Convert LLMEngineOutput to PyObject
                     let py_response = Python::with_gil(|py| {
                         pythonize(py, &response.data)
@@ -976,6 +992,11 @@ impl KvPushRouter {
                         }
                     }
                 }
+
+                // Observe per-worker ITL gauge at stream end
+                if let Some(ref tracker) = tracker {
+                    tracker.observe_finish_gauges();
+                }
             });
 
             // Return a Python async generator wrapper

lib/llm/src/discovery.rs

@@ -12,7 +12,5 @@ pub use watcher::{ModelUpdate, ModelWatcher};
 
 mod worker_monitor;
 pub use worker_monitor::{
-    KvWorkerMonitor, LoadThresholdConfig, WORKER_ACTIVE_DECODE_BLOCKS_GAUGE,
-    WORKER_ACTIVE_PREFILL_TOKENS_GAUGE, WORKER_TYPE_DECODE, WORKER_TYPE_PREFILL, WorkerLoadState,
-    register_worker_load_metrics,
+    KvWorkerMonitor, LoadThresholdConfig, WORKER_TYPE_DECODE, WORKER_TYPE_PREFILL, WorkerLoadState,
 };

lib/llm/src/discovery/worker_monitor.rs

@@ -3,13 +3,12 @@
 
 use std::collections::HashMap;
 use std::sync::Arc;
+use std::sync::RwLock;
 use std::sync::atomic::{AtomicBool, AtomicU32, AtomicU64, Ordering};
-use std::sync::{LazyLock, RwLock};
 
 use tokio::sync::Notify;
 
 use dashmap::DashMap;
-use prometheus::{IntGaugeVec, Opts, Registry};
 use serde::{Deserialize, Serialize};
 
 use crate::http::service::metrics::{
@@ -17,11 +16,11 @@ use crate::http::service::metrics::{
     WORKER_LAST_TIME_TO_FIRST_TOKEN_GAUGE,
 };
 use crate::kv_router::KV_METRICS_SUBJECT;
+use crate::kv_router::metrics::WORKER_LOAD_METRICS;
 use crate::kv_router::protocols::ActiveLoad;
 use crate::model_card::ModelDeploymentCard;
 use dynamo_runtime::component::Client;
 use dynamo_runtime::discovery::{DiscoveryQuery, watch_and_extract_field};
-use dynamo_runtime::metrics::prometheus_names::frontend_service;
 use dynamo_runtime::pipeline::{WorkerLoadMonitor, async_trait};
 use dynamo_runtime::traits::DistributedRuntimeProvider;
 use dynamo_runtime::transports::event_plane::EventSubscriber;
@@ -29,62 +28,18 @@ use dynamo_runtime::transports::event_plane::EventSubscriber;
 // Re-export worker type constants from timing.rs (single source of truth)
 pub use crate::protocols::common::timing::{WORKER_TYPE_DECODE, WORKER_TYPE_PREFILL};
 
-/// Global Prometheus gauge for active decode blocks per worker (labels: worker_id, dp_rank, worker_type)
-/// This is shared across all KvWorkerMonitor instances.
-pub static WORKER_ACTIVE_DECODE_BLOCKS_GAUGE: LazyLock<IntGaugeVec> = LazyLock::new(|| {
-    IntGaugeVec::new(
-        Opts::new(
-            format!(
-                "dynamo_frontend_{}",
-                frontend_service::WORKER_ACTIVE_DECODE_BLOCKS
-            ),
-            "Active KV cache decode blocks per worker",
-        ),
-        &["worker_id", "dp_rank", "worker_type"],
-    )
-    .expect("Failed to create worker_active_decode_blocks gauge")
-});
-
-/// Global Prometheus gauge for active prefill tokens per worker (labels: worker_id, dp_rank, worker_type)
-/// This is shared across all KvWorkerMonitor instances.
-pub static WORKER_ACTIVE_PREFILL_TOKENS_GAUGE: LazyLock<IntGaugeVec> = LazyLock::new(|| {
-    IntGaugeVec::new(
-        Opts::new(
-            format!(
-                "dynamo_frontend_{}",
-                frontend_service::WORKER_ACTIVE_PREFILL_TOKENS
-            ),
-            "Active prefill tokens queued per worker",
-        ),
-        &["worker_id", "dp_rank", "worker_type"],
-    )
-    .expect("Failed to create worker_active_prefill_tokens gauge")
-});
-
-/// Register the global worker load Prometheus metrics with the given registry.
-///
-/// This should be called once during HTTP service setup to expose the worker load
-/// metrics via the `/metrics` endpoint.
-///
-/// # Errors
-/// Returns an error if the metrics are already registered with the registry.
-pub fn register_worker_load_metrics(registry: &Registry) -> Result<(), prometheus::Error> {
-    registry.register(Box::new(WORKER_ACTIVE_DECODE_BLOCKS_GAUGE.clone()))?;
-    registry.register(Box::new(WORKER_ACTIVE_PREFILL_TOKENS_GAUGE.clone()))?;
-    Ok(())
-}
-
 /// Clean up all Prometheus metrics for a worker across the specified dp_ranks.
 ///
 /// This removes metrics with the given worker_id, dp_rank, and worker_type label combination.
 /// Called when workers are removed to prevent stale metrics from accumulating.
 fn cleanup_worker_metrics(worker_id: u64, dp_ranks: &[u32], worker_type: &str) {
     let worker_id_str = worker_id.to_string();
+    let m = &*WORKER_LOAD_METRICS;
     for dp_rank in dp_ranks {
         let dp_rank_str = dp_rank.to_string();
         let labels = &[worker_id_str.as_str(), dp_rank_str.as_str(), worker_type];
-        let _ = WORKER_ACTIVE_DECODE_BLOCKS_GAUGE.remove_label_values(labels);
-        let _ = WORKER_ACTIVE_PREFILL_TOKENS_GAUGE.remove_label_values(labels);
+        let _ = m.active_decode_blocks.remove_label_values(labels);
+        let _ = m.active_prefill_tokens.remove_label_values(labels);
         let _ = WORKER_LAST_TIME_TO_FIRST_TOKEN_GAUGE.remove_label_values(labels);
         let _ = WORKER_LAST_INPUT_SEQUENCE_TOKENS_GAUGE.remove_label_values(labels);
         let _ = WORKER_LAST_INTER_TOKEN_LATENCY_GAUGE.remove_label_values(labels);
@@ -215,9 +170,9 @@ impl WorkerLoadState {
 /// Cloning shares state via internal Arc-wrapped fields. This allows multiple pipelines
 /// (e.g., chat and completions) to share the same monitor instance.
 ///
-/// Prometheus metrics are exposed via the global gauges [`WORKER_ACTIVE_DECODE_BLOCKS_GAUGE`]
-/// and [`WORKER_ACTIVE_PREFILL_TOKENS_GAUGE`], which should be registered with the HTTP
-/// service's Prometheus registry using [`register_worker_load_metrics`].
+/// Prometheus metrics are exposed via [`WORKER_LOAD_METRICS`] (defined in `kv_router::sequence`),
+/// which should be registered with the HTTP service's Prometheus registry using
+/// [`register_worker_load_metrics`](crate::kv_router::metrics::register_worker_load_metrics).
 ///
 /// In disaggregated mode, use `set_prefill_client` to register the prefill endpoint for
 /// proper TTFT metric cleanup when prefill workers are removed.
@@ -251,8 +206,9 @@ impl KvWorkerMonitor {
     /// - `active_prefill_tokens_threshold`: DEFAULT_MAX_TOKENS (effectively disabled)
     /// - `active_prefill_tokens_threshold_frac`: 1.5 (effectively disabled)
     ///
-    /// Prometheus metrics are exposed via the global gauges and should be registered
-    /// using [`register_worker_load_metrics`] during HTTP service setup.
+    /// Prometheus metrics are exposed via [`WORKER_LOAD_METRICS`] and should be registered
+    /// using [`register_worker_load_metrics`](crate::kv_router::metrics::register_worker_load_metrics)
+    /// during HTTP service setup.
     ///
     /// For disaggregated mode, call `set_prefill_client` after creation to enable
     /// proper TTFT metric cleanup when prefill workers are removed.

lib/llm/src/http/service/metrics.rs

@@ -111,7 +111,7 @@ pub fn register_worker_timing_metrics(registry: &Registry) -> Result<(), prometh
 /// # Note
 /// With 2 significant figures, there are roughly 90 unique values per order of magnitude.
 /// Requesting more buckets than can be uniquely represented will result in deduplication.
-fn generate_log_buckets(min: f64, max: f64, count: usize) -> Vec<f64> {
+pub fn generate_log_buckets(min: f64, max: f64, count: usize) -> Vec<f64> {
     if count == 0 {
         return vec![];
     }
@@ -153,7 +153,7 @@ fn generate_log_buckets(min: f64, max: f64, count: usize) -> Vec<f64> {
 }
 
 /// Round a number to a specified number of significant figures
-fn round_to_sig_figs(value: f64, sig_figs: u32) -> f64 {
+pub fn round_to_sig_figs(value: f64, sig_figs: u32) -> f64 {
     if value == 0.0 {
         return 0.0;
     }

lib/llm/src/http/service/service_v2.rs

@@ -16,8 +16,9 @@ use super::Metrics;
 use super::RouteDoc;
 use super::metrics;
 use super::metrics::register_worker_timing_metrics;
-use crate::discovery::{ModelManager, register_worker_load_metrics};
+use crate::discovery::ModelManager;
 use crate::endpoint_type::EndpointType;
+use crate::kv_router::metrics::{register_routing_overhead_metrics, register_worker_load_metrics};
 use crate::request_template::RequestTemplate;
 use anyhow::Result;
 use axum_server::tls_rustls::RustlsConfig;
@@ -391,6 +392,12 @@ impl HttpServiceConfigBuilder {
             tracing::warn!("Failed to register worker timing metrics: {}", e);
         }
 
+        // Register routing overhead metrics (block hashing, find matches, scheduling latencies)
+        // These are updated by KvRouter::find_best_match on every routing decision
+        if let Err(e) = register_routing_overhead_metrics(&registry) {
+            tracing::warn!("Failed to register routing overhead metrics: {}", e);
+        }
+
         let mut router = axum::Router::new();
 
         let mut all_docs = Vec::new();

lib/llm/src/kv_router.rs

@@ -2,9 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 
 use std::collections::HashMap;
-use std::time::Duration;
-#[cfg(feature = "bench")]
-use std::time::Instant;
+use std::time::{Duration, Instant};
 
 use anyhow::Result;
 use dynamo_runtime::{
@@ -27,6 +25,7 @@ pub use dynamo_kv_router::indexer;
 pub use dynamo_kv_router::protocols;
 
 pub mod config;
+pub mod metrics;
 pub mod prefill_router;
 pub mod publisher;
 pub mod push_router;
@@ -63,7 +62,6 @@ pub const KV_METRICS_ENDPOINT: &str = "load_metrics";
 
 // for metric publishing (push-based)
 pub const KV_EVENT_SUBJECT: &str = "kv-events";
-pub const KV_HIT_RATE_SUBJECT: &str = "kv-hit-rate";
 pub const KV_METRICS_SUBJECT: &str = "kv_metrics";
 
 // for inter-router comms
@@ -316,7 +314,6 @@ impl KvRouter {
         update_states: bool,
         lora_name: Option<String>,
     ) -> anyhow::Result<(WorkerWithDpRank, u32)> {
-        #[cfg(feature = "bench")]
         let start = Instant::now();
 
         if update_states && context_id.is_none() {
@@ -326,16 +323,16 @@ impl KvRouter {
         let isl_tokens = tokens.len();
 
         let block_hashes = compute_block_hash_for_seq(tokens, self.block_size, None);
-        #[cfg(feature = "bench")]
         let hash_elapsed = start.elapsed();
+
         let overlap_scores = self.indexer.find_matches(block_hashes).await?;
-        #[cfg(feature = "bench")]
         let find_matches_elapsed = start.elapsed();
 
         // Compute seq_hashes only if scheduler needs it for active blocks tracking
         let maybe_seq_hashes = self
             .kv_router_config
             .compute_seq_hashes_for_tracking(tokens, self.block_size);
+        let seq_hash_elapsed = start.elapsed();
 
         let best_worker = self
             .scheduler
@@ -349,19 +346,25 @@ impl KvRouter {
                 lora_name,
             )
             .await?;
+        let total_elapsed = start.elapsed();
+
+        metrics::ROUTING_OVERHEAD_METRICS.observe(
+            hash_elapsed,
+            find_matches_elapsed,
+            seq_hash_elapsed,
+            total_elapsed,
+        );
 
         #[cfg(feature = "bench")]
-        {
-            let total_elapsed = start.elapsed();
-            tracing::info!(
-                isl_tokens,
-                hash_us = hash_elapsed.as_micros() as u64,
-                find_matches_us = (find_matches_elapsed - hash_elapsed).as_micros() as u64,
-                schedule_us = (total_elapsed - find_matches_elapsed).as_micros() as u64,
-                total_us = total_elapsed.as_micros() as u64,
-                "find_best_match completed"
-            );
-        }
+        tracing::info!(
+            isl_tokens,
+            hash_us = hash_elapsed.as_micros() as u64,
+            find_matches_us = (find_matches_elapsed - hash_elapsed).as_micros() as u64,
+            seq_hash_us = (seq_hash_elapsed - find_matches_elapsed).as_micros() as u64,
+            schedule_us = (total_elapsed - seq_hash_elapsed).as_micros() as u64,
+            total_us = total_elapsed.as_micros() as u64,
+            "find_best_match completed"
+        );
 
         // Note: Routing decision recording (for approximate mode) is now handled
         // by KvPushRouter::generate after select_worker returns.

lib/llm/src/kv_router/metrics.rs

+// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+//! Prometheus metrics for the KV router.
+//!
+//! This module centralizes all router-side Prometheus metric definitions:
+//! - [`WorkerLoadMetrics`]: Per-worker active decode blocks and prefill tokens gauges.
+//! - [`RoutingOverheadMetrics`]: Per-request routing phase latency histograms.
+
+use std::sync::{Arc, LazyLock, OnceLock};
+use std::time::Duration;
+
+use dynamo_runtime::component::Component;
+use dynamo_runtime::metrics::MetricsHierarchy;
+use dynamo_runtime::metrics::prometheus_names::{
+    frontend_service, labels, name_prefix, routing_overhead,
+};
+use prometheus::{IntGaugeVec, Opts};
+
+use crate::http::service::metrics::generate_log_buckets;
+
+// ---------------------------------------------------------------------------
+// Worker load metrics (gauges)
+// ---------------------------------------------------------------------------
+
+/// Per-worker active load gauges, published by `ActiveSequencesMultiWorker`
+/// and cleaned up by `KvWorkerMonitor` when workers disappear.
+pub struct WorkerLoadMetrics {
+    pub active_decode_blocks: IntGaugeVec,
+    pub active_prefill_tokens: IntGaugeVec,
+}
+
+impl WorkerLoadMetrics {
+    pub fn observe(
+        &self,
+        worker_id: u64,
+        dp_rank: u32,
+        worker_type: &str,
+        active_blocks: usize,
+        active_tokens: usize,
+    ) {
+        let worker_id_str = worker_id.to_string();
+        let dp_rank_str = dp_rank.to_string();
+        let labels = &[worker_id_str.as_str(), dp_rank_str.as_str(), worker_type];
+        self.active_decode_blocks
+            .with_label_values(labels)
+            .set(active_blocks as i64);
+        self.active_prefill_tokens
+            .with_label_values(labels)
+            .set(active_tokens as i64);
+    }
+}
+
+pub static WORKER_LOAD_METRICS: LazyLock<WorkerLoadMetrics> = LazyLock::new(|| WorkerLoadMetrics {
+    active_decode_blocks: IntGaugeVec::new(
+        Opts::new(
+            format!(
+                "{}_{}",
+                name_prefix::FRONTEND,
+                frontend_service::WORKER_ACTIVE_DECODE_BLOCKS
+            ),
+            "Active KV cache decode blocks per worker",
+        ),
+        &[labels::WORKER_ID, labels::DP_RANK, labels::WORKER_TYPE],
+    )
+    .expect("Failed to create worker_active_decode_blocks gauge"),
+    active_prefill_tokens: IntGaugeVec::new(
+        Opts::new(
+            format!(
+                "{}_{}",
+                name_prefix::FRONTEND,
+                frontend_service::WORKER_ACTIVE_PREFILL_TOKENS
+            ),
+            "Active prefill tokens queued per worker",
+        ),
+        &[labels::WORKER_ID, labels::DP_RANK, labels::WORKER_TYPE],
+    )
+    .expect("Failed to create worker_active_prefill_tokens gauge"),
+});
+
+/// Register the worker load gauges with the given Prometheus registry.
+pub fn register_worker_load_metrics(
+    registry: &prometheus::Registry,
+) -> Result<(), prometheus::Error> {
+    let m = &*WORKER_LOAD_METRICS;
+    registry.register(Box::new(m.active_decode_blocks.clone()))?;
+    registry.register(Box::new(m.active_prefill_tokens.clone()))?;
+    Ok(())
+}
+
+// ---------------------------------------------------------------------------
+// Routing overhead metrics (histograms)
+// ---------------------------------------------------------------------------
+
+/// Per-request routing phase latency histograms (milliseconds).
+pub struct RoutingOverheadMetrics {
+    pub block_hashing: prometheus::Histogram,
+    pub indexer_find_matches: prometheus::Histogram,
+    pub seq_hashing: prometheus::Histogram,
+    pub scheduling: prometheus::Histogram,
+    pub total: prometheus::Histogram,
+}
+
+pub static ROUTING_OVERHEAD_METRICS: LazyLock<RoutingOverheadMetrics> = LazyLock::new(|| {
+    // Buckets from 0.0001ms (0.1μs) to ~10ms, exponential with factor 2
+    let buckets = prometheus::exponential_buckets(0.0001, 2.0, 18)
+        .expect("exponential buckets should not fail");
+    let make = |suffix: &str, help: &str| {
+        let name = format!("{}_{}", name_prefix::ROUTING_OVERHEAD, suffix);
+        prometheus::Histogram::with_opts(
+            prometheus::HistogramOpts::new(name, help).buckets(buckets.clone()),
+        )
+        .expect("histogram creation should not fail")
+    };
+    RoutingOverheadMetrics {
+        block_hashing: make(
+            routing_overhead::BLOCK_HASHING_MS,
+            "Time spent computing block hashes in milliseconds",
+        ),
+        indexer_find_matches: make(
+            routing_overhead::INDEXER_FIND_MATCHES_MS,
+            "Time spent in indexer find_matches in milliseconds",
+        ),
+        seq_hashing: make(
+            routing_overhead::SEQ_HASHING_MS,
+            "Time spent computing sequence hashes in milliseconds",
+        ),
+        scheduling: make(
+            routing_overhead::SCHEDULING_MS,
+            "Time spent in scheduler worker selection in milliseconds",
+        ),
+        total: make(
+            routing_overhead::TOTAL_MS,
+            "Total routing overhead per request in milliseconds",
+        ),
+    }
+});
+
+impl RoutingOverheadMetrics {
+    /// Observe routing overhead timings in milliseconds.
+    pub fn observe(
+        &self,
+        hash_elapsed: Duration,
+        find_matches_elapsed: Duration,
+        seq_hash_elapsed: Duration,
+        total_elapsed: Duration,
+    ) {
+        self.block_hashing
+            .observe(hash_elapsed.as_secs_f64() * 1000.0);
+        self.indexer_find_matches.observe(
+            find_matches_elapsed
+                .saturating_sub(hash_elapsed)
+                .as_secs_f64()
+                * 1000.0,
+        );
+        self.seq_hashing.observe(
+            seq_hash_elapsed
+                .saturating_sub(find_matches_elapsed)
+                .as_secs_f64()
+                * 1000.0,
+        );
+        self.scheduling
+            .observe(total_elapsed.saturating_sub(seq_hash_elapsed).as_secs_f64() * 1000.0);
+        self.total.observe(total_elapsed.as_secs_f64() * 1000.0);
+    }
+}
+
+/// Register the routing overhead histograms with the given Prometheus registry.
+pub fn register_routing_overhead_metrics(
+    registry: &prometheus::Registry,
+) -> Result<(), prometheus::Error> {
+    let m = &*ROUTING_OVERHEAD_METRICS;
+    registry.register(Box::new(m.block_hashing.clone()))?;
+    registry.register(Box::new(m.indexer_find_matches.clone()))?;
+    registry.register(Box::new(m.seq_hashing.clone()))?;
+    registry.register(Box::new(m.scheduling.clone()))?;
+    registry.register(Box::new(m.total.clone()))?;
+    Ok(())
+}
+
+// ---------------------------------------------------------------------------
+// Router request metrics (component-scoped aggregate histograms + counter)
+// ---------------------------------------------------------------------------
+
+/// Aggregate per-request metrics observed at the router level.
+/// Component-scoped via `from_component()` to get automatic namespace/component labels.
+/// Uses `router_` prefix to distinguish from `ResponseMetricCollector` metrics.
+pub struct RouterRequestMetrics {
+    pub requests_total: prometheus::IntCounter,
+    pub time_to_first_token_seconds: prometheus::Histogram,
+    pub inter_token_latency_seconds: prometheus::Histogram,
+    pub input_sequence_tokens: prometheus::Histogram,
+    pub output_sequence_tokens: prometheus::Histogram,
+}
+
+static ROUTER_REQUEST_METRICS: OnceLock<Arc<RouterRequestMetrics>> = OnceLock::new();
+
+impl RouterRequestMetrics {
+    fn new(
+        requests_total: prometheus::IntCounter,
+        time_to_first_token_seconds: prometheus::Histogram,
+        inter_token_latency_seconds: prometheus::Histogram,
+        input_sequence_tokens: prometheus::Histogram,
+        output_sequence_tokens: prometheus::Histogram,
+    ) -> Self {
+        Self {
+            requests_total,
+            time_to_first_token_seconds,
+            inter_token_latency_seconds,
+            input_sequence_tokens,
+            output_sequence_tokens,
+        }
+    }
+
+    /// Create from a Component, memoized in a static OnceLock.
+    pub fn from_component(component: &Component) -> Arc<Self> {
+        ROUTER_REQUEST_METRICS
+            .get_or_init(|| {
+                let metrics = component.metrics();
+                let requests_total = metrics
+                    .create_intcounter(
+                        "router_requests_total",
+                        "Total number of requests processed by the router",
+                        &[],
+                    )
+                    .expect("failed to create router_requests_total");
+                let time_to_first_token_seconds = metrics
+                    .create_histogram(
+                        "router_time_to_first_token_seconds",
+                        "Time to first token observed at the router",
+                        &[],
+                        Some(generate_log_buckets(0.001, 480.0, 18)),
+                    )
+                    .expect("failed to create router_time_to_first_token_seconds");
+                let inter_token_latency_seconds = metrics
+                    .create_histogram(
+                        "router_inter_token_latency_seconds",
+                        "Average inter-token latency observed at the router",
+                        &[],
+                        Some(generate_log_buckets(0.001, 2.0, 13)),
+                    )
+                    .expect("failed to create router_inter_token_latency_seconds");
+                let input_sequence_tokens = metrics
+                    .create_histogram(
+                        "router_input_sequence_tokens",
+                        "Input sequence length in tokens observed at the router",
+                        &[],
+                        Some(generate_log_buckets(50.0, 128000.0, 12)),
+                    )
+                    .expect("failed to create router_input_sequence_tokens");
+                let output_sequence_tokens = metrics
+                    .create_histogram(
+                        "router_output_sequence_tokens",
+                        "Output sequence length in tokens observed at the router",
+                        &[],
+                        Some(generate_log_buckets(50.0, 32000.0, 10)),
+                    )
+                    .expect("failed to create router_output_sequence_tokens");
+                Arc::new(Self::new(
+                    requests_total,
+                    time_to_first_token_seconds,
+                    inter_token_latency_seconds,
+                    input_sequence_tokens,
+                    output_sequence_tokens,
+                ))
+            })
+            .clone()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use prometheus::{Encoder, TextEncoder};
+
+    fn gather_pef(registry: &prometheus::Registry) -> String {
+        let encoder = TextEncoder::new();
+        let mut buffer = Vec::new();
+        encoder.encode(&registry.gather(), &mut buffer).unwrap();
+        String::from_utf8(buffer).unwrap()
+    }
+
+    #[test]
+    fn test_worker_load_metrics_pef() {
+        let registry = prometheus::Registry::new();
+        let metrics = WorkerLoadMetrics {
+            active_decode_blocks: IntGaugeVec::new(
+                Opts::new(
+                    format!(
+                        "{}_{}",
+                        name_prefix::FRONTEND,
+                        frontend_service::WORKER_ACTIVE_DECODE_BLOCKS
+                    ),
+                    "Active KV cache decode blocks per worker",
+                ),
+                &[labels::WORKER_ID, labels::DP_RANK, labels::WORKER_TYPE],
+            )
+            .unwrap(),
+            active_prefill_tokens: IntGaugeVec::new(
+                Opts::new(
+                    format!(
+                        "{}_{}",
+                        name_prefix::FRONTEND,
+                        frontend_service::WORKER_ACTIVE_PREFILL_TOKENS
+                    ),
+                    "Active prefill tokens queued per worker",
+                ),
+                &[labels::WORKER_ID, labels::DP_RANK, labels::WORKER_TYPE],
+            )
+            .unwrap(),
+        };
+        registry
+            .register(Box::new(metrics.active_decode_blocks.clone()))
+            .unwrap();
+        registry
+            .register(Box::new(metrics.active_prefill_tokens.clone()))
+            .unwrap();
+
+        metrics.observe(123, 0, "decode", 42, 100);
+
+        let output = gather_pef(&registry);
+        let expected = "\
+# HELP dynamo_frontend_worker_active_decode_blocks Active KV cache decode blocks per worker
+# TYPE dynamo_frontend_worker_active_decode_blocks gauge
+dynamo_frontend_worker_active_decode_blocks{dp_rank=\"0\",worker_id=\"123\",worker_type=\"decode\"} 42
+# HELP dynamo_frontend_worker_active_prefill_tokens Active prefill tokens queued per worker
+# TYPE dynamo_frontend_worker_active_prefill_tokens gauge
+dynamo_frontend_worker_active_prefill_tokens{dp_rank=\"0\",worker_id=\"123\",worker_type=\"decode\"} 100
+";
+        assert_eq!(
+            output, expected,
+            "\nActual PEF:\n{output}\nExpected PEF:\n{expected}"
+        );
+    }
+
+    #[test]
+    fn test_routing_overhead_metric_names_pef() {
+        let registry = prometheus::Registry::new();
+        let buckets = prometheus::exponential_buckets(0.0001, 2.0, 18).unwrap();
+        let make = |suffix: &str, help: &str| {
+            let name = format!("{}_{}", name_prefix::ROUTING_OVERHEAD, suffix);
+            prometheus::Histogram::with_opts(
+                prometheus::HistogramOpts::new(name, help).buckets(buckets.clone()),
+            )
+            .unwrap()
+        };
+
+        let total = make(
+            routing_overhead::TOTAL_MS,
+            "Total routing overhead per request in milliseconds",
+        );
+        registry.register(Box::new(total.clone())).unwrap();
+        total.observe(1.5);
+
+        let output = gather_pef(&registry);
+        assert!(
+            output.contains("# HELP dynamo_routing_overhead_total_ms"),
+            "PEF missing HELP for routing overhead metric"
+        );
+        assert!(
+            output.contains("# TYPE dynamo_routing_overhead_total_ms histogram"),
+            "PEF missing TYPE for routing overhead metric"
+        );
+        assert!(
+            output.contains("dynamo_routing_overhead_total_ms_count 1"),
+            "PEF missing observation count"
+        );
+    }
+
+    #[test]
+    fn test_routing_overhead_saturating_sub() {
+        let buckets = prometheus::exponential_buckets(0.0001, 2.0, 18).unwrap();
+        let make = |name: &str| {
+            prometheus::Histogram::with_opts(
+                prometheus::HistogramOpts::new(name, "test").buckets(buckets.clone()),
+            )
+            .unwrap()
+        };
+        let metrics = RoutingOverheadMetrics {
+            block_hashing: make("test_block_hashing_ms"),
+            indexer_find_matches: make("test_find_matches_ms"),
+            seq_hashing: make("test_seq_hashing_ms"),
+            scheduling: make("test_scheduling_ms"),
+            total: make("test_total_ms"),
+        };
+
+        // Out-of-order durations: each phase < previous (would panic without saturating_sub)
+        metrics.observe(
+            Duration::from_millis(10),
+            Duration::from_millis(5),
+            Duration::from_millis(3),
+            Duration::from_millis(1),
+        );
+        // Reaching here without panic confirms saturating_sub works
+    }
+}

lib/llm/src/kv_router/prefill_router.rs

@@ -348,7 +348,7 @@ impl PrefillRouter {
     ///
     /// If `phase_permit` is provided, it is dropped after the first output is received,
     /// allowing subsequent `set_phase` calls to proceed. This is used in the bootstrap
-    /// optimization path to ensure `record_worker` completes before the phase changes.
+    /// optimization path to ensure `record_worker_full` completes before the phase changes.
     ///
     /// Returns (PrefillResult, Option<(worker_id, dp_rank)>).
     async fn execute_prefill(
@@ -363,7 +363,7 @@ impl PrefillRouter {
             .await
             .map_err(|e| PrefillError::PrefillError(e.to_string()))?;
 
-        // Drop phase permit now - routing is complete, record_worker was called in select_worker.
+        // Drop phase permit now - routing is complete, record_worker_full was called in select_worker.
         // This unblocks set_phase(Decode) in the main task without waiting for prefill output.
         drop(phase_permit);
 
@@ -530,7 +530,6 @@ impl
         }
         let tracker = req.tracker.as_ref().unwrap();
         let prefill_phase_permit = tracker.set_phase(RequestPhase::Prefill).await;
-        tracker.record_prefill_start();
 
         // Prepare prefill request with max_tokens = 1 (clone after tracker is set)
         let mut prefill_req = req.clone();
@@ -557,14 +556,6 @@ impl
                     router.select_next_worker();
                 }
 
-                // Record prefill worker on the main request's tracker for metrics.
-                // (The cloned prefill_req has its own tracker, so we need to record here)
-                // Worker type is stored at routing time to avoid expensive MDC lookups when
-                // updating Prometheus TTFT metrics later in the response stream.
-                if let Some(ref tracker) = req.tracker {
-                    tracker.record_prefill_worker_full(worker_id, dp_rank, WORKER_TYPE_PREFILL);
-                }
-
                 let routing = prefill_req.routing_mut();
                 routing.prefill_worker_id = Some(worker_id);
                 routing.dp_rank = Some(dp_rank);
@@ -573,7 +564,7 @@ impl
                 let prefill_context = Context::with_id(prefill_req, request_id.clone());
                 engine_ctx.link_child(prefill_context.context());
 
-                // Pass phase permit to spawned task - it drops after first output (record_worker complete)
+                // Pass phase permit to spawned task - it drops after first output (record_worker_full complete)
                 // This allows set_phase(Decode) below to proceed only after prefill routing is done
                 self.spawn_prefill_task(prefill_context, Some(worker_id), prefill_phase_permit);
 
@@ -591,16 +582,6 @@ impl
 
                 let result = self.call_prefill(prefill_context).await;
 
-                // Record prefill worker on the main request's tracker for metrics.
-                // (call_prefill returns the worker_id and dp_rank from the prefill routing)
-                // Worker type is stored at routing time to avoid expensive MDC lookups when
-                // updating Prometheus TTFT metrics later in the response stream.
-                if let Ok((_, Some((worker_id, dp_rank)))) = &result
-                    && let Some(ref tracker) = req.tracker
-                {
-                    tracker.record_prefill_worker_full(*worker_id, *dp_rank, WORKER_TYPE_PREFILL);
-                }
-
                 result.map(|(result, worker_info)| {
                     (Some(result), worker_info.map(|(id, _)| id), None)
                 })
@@ -625,7 +606,7 @@ impl
 
                 // Set phase to Decode for the decode request.
                 // In bootstrap path, this blocks until the spawned prefill task drops its permit
-                // (after first output / record_worker completes), ensuring correct phase for routing.
+                // (after first output / record_worker_full completes), ensuring correct phase for routing.
                 if let Some(ref tracker) = req.tracker {
                     let _decode_permit = tracker.set_phase(RequestPhase::Decode).await;
                     // Permit is dropped immediately - decode proceeds, no need to hold it

lib/llm/src/kv_router/push_router.rs

@@ -17,6 +17,7 @@ use serde_json::json;
 use crate::{
     kv_router::{
         KvRouter,
+        metrics::RouterRequestMetrics,
         protocols::{TokensWithHashes, WorkerWithDpRank},
     },
     preprocessor::PreprocessedRequest,
@@ -219,20 +220,25 @@ impl AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<LLMEngineOutpu
             }
         }
 
-        // Record metrics in tracker: KV hit rate, worker ID, and worker type based on phase.
-        // Worker type is stored at routing time to avoid expensive MDC lookups when
-        // updating Prometheus metrics (TTFT/ITL) later in the response stream.
+        // Record routing metrics on tracker and observe ISL + prefill start.
+        let request_metrics =
+            RouterRequestMetrics::from_component(self.chooser.client().endpoint.component());
         if let Some(ref tracker) = request.tracker {
             let isl_blocks = request.token_ids.len().div_ceil(block_size);
             tracker.record_kv_hit(overlap_amount, isl_blocks);
+            tracker.record_isl(
+                request.token_ids.len(),
+                overlap_amount as usize * block_size,
+            );
             tracker.record_worker_full(instance_id, dp_rank, self.chooser.worker_type());
         }
+        request_metrics
+            .input_sequence_tokens
+            .observe(request.token_ids.len() as f64);
 
         // Handle query-only requests: early return with worker info
         if is_query_only {
             let stream_context = request.context().clone();
-            // Tracker is always created for query-only requests (delta generator enables tracking
-            // when query_instance_id annotation is present)
             let worker_id_info = request.tracker.as_ref().and_then(|t| t.get_worker_info());
 
             tracing::trace!(
@@ -262,11 +268,17 @@ impl AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<LLMEngineOutpu
             .and_then(|r| r.expected_output_tokens);
         let track_output_blocks =
             self.chooser.kv_router_config().router_track_output_blocks && handle_local_updates;
+        let tracker = request.tracker.clone();
 
         let (mut backend_input, context) = request.into_parts();
         backend_input.routing_mut().dp_rank = Some(dp_rank);
         let updated_request = context.map(|_| backend_input);
 
+        // Record prefill start right before pushing to backend (OnceLock: first call wins).
+        if let Some(ref tracker) = tracker {
+            tracker.record_prefill_start();
+        }
+
         let chooser = self.chooser.clone();
         let mut response_stream = self.inner.direct(updated_request, instance_id).await?;
         let stream_context = response_stream.context();
@@ -277,6 +289,7 @@ impl AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<LLMEngineOutpu
         // When false, an external caller (e.g., GAIE sidecar) handles bookkeeping via C FFI.
         let wrapped_stream = Box::pin(async_stream::stream! {
             let mut prefill_marked = false;
+            let mut first_token_recorded = false;
 
             // Output block tracking state
             let mut cumulative_osl: usize = 0;
@@ -310,13 +323,27 @@ impl AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<LLMEngineOutpu
                             }
                         }
 
+                        let new_tokens = item.data.as_ref()
+                            .map(|d| d.token_ids.len())
+                            .unwrap_or(0);
+
+                        // Record first token time on tracker when actual tokens arrive
+                        if !first_token_recorded && new_tokens > 0 {
+                            if let Some(ref tracker) = tracker {
+                                tracker.record_first_token();
+                                if let Some(ttft) = tracker.ttft_ms() {
+                                    request_metrics
+                                        .time_to_first_token_seconds
+                                        .observe(ttft / 1000.0);
+                                }
+                            }
+                            first_token_recorded = true;
+                        }
+
+                        cumulative_osl += new_tokens;
+
                         // Track output blocks if enabled
                         if track_output_blocks {
-                            let new_tokens = item.data.as_ref()
-                                .map(|d| d.token_ids.len())
-                                .unwrap_or(0);
-                            cumulative_osl += new_tokens;
-
                             let new_total_blocks = (isl_tokens + cumulative_osl).div_ceil(block_size);
                             if new_total_blocks > current_total_blocks {
                                 // New block boundary crossed - add output block with decay
@@ -329,6 +356,18 @@ impl AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<LLMEngineOutpu
                                         "Failed to add output block for request {context_id}: {e}"
                                     );
                                 }
+
+                                // Update tracker and observe avg ITL at each block boundary
+                                if let Some(ref tracker) = tracker {
+                                    tracker.record_osl(cumulative_osl);
+                                    tracker.record_finish();
+                                    if let Some(avg_itl) = tracker.avg_itl_ms() {
+                                        request_metrics
+                                            .inter_token_latency_seconds
+                                            .observe(avg_itl / 1000.0);
+                                    }
+                                }
+
                                 current_total_blocks = new_total_blocks;
                             }
                         }
@@ -338,6 +377,17 @@ impl AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<LLMEngineOutpu
                 }
             }
 
+            // Record final aggregate metrics (histograms sampled once per request)
+            if let Some(ref tracker) = tracker {
+                tracker.record_finish();
+                tracker.record_osl(cumulative_osl);
+
+                request_metrics
+                    .output_sequence_tokens
+                    .observe(cumulative_osl as f64);
+            }
+            request_metrics.requests_total.inc();
+
             // Only call free() if we handle local updates.
             // When handle_local_updates=false, external caller handles cleanup via C FFI.
             if handle_local_updates

lib/llm/src/kv_router/scheduler.rs

@@ -6,7 +6,6 @@ use crate::local_model::runtime_config::ModelRuntimeConfig;
 use anyhow::Result;
 use dynamo_runtime::component::Component;
 use dynamo_runtime::traits::DistributedRuntimeProvider;
-use dynamo_runtime::transports::event_plane::EventPublisher;
 use rand::Rng;
 use serde::{Deserialize, Serialize};
 use std::collections::{HashMap, HashSet};
@@ -15,7 +14,6 @@ use std::time::Duration;
 #[cfg(feature = "bench")]
 use std::time::Instant;
 
-use super::KV_HIT_RATE_SUBJECT;
 use super::KvRouterConfig;
 use super::RouterConfigOverride;
 use super::WorkerSelector;
@@ -24,15 +22,6 @@ use super::sequence::{ActiveSequencesMultiWorker, SequenceError};
 
 use dynamo_tokens::SequenceHash;
 
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct KVHitRateEvent {
-    pub worker_id: WorkerId,
-    #[serde(default)]
-    pub dp_rank: DpRank,
-    pub isl_blocks: usize,
-    pub overlap_blocks: u32,
-}
-
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct PotentialLoad {
     pub worker_id: WorkerId,
@@ -160,10 +149,6 @@ impl KvScheduler {
         let scheduler_rx = workers_with_configs.clone();
         let (request_tx, request_rx) = tokio::sync::mpsc::channel::<SchedulingRequest>(1024);
         let scheduler_cancel_token = component.drt().primary_token();
-        let hit_rate_publisher =
-            EventPublisher::for_namespace(component.namespace(), KV_HIT_RATE_SUBJECT)
-                .await
-                .map_err(|e| KvSchedulerError::InitFailed(e.to_string()))?;
 
         // Background task to handle scheduling requests
         tokio::spawn(async move {
@@ -199,16 +184,6 @@ impl KvScheduler {
 
                 match selector.select_worker(&workers, &request, block_size) {
                     Ok(selection) => {
-                        let event = KVHitRateEvent {
-                            worker_id: selection.worker.worker_id,
-                            dp_rank: selection.worker.dp_rank,
-                            isl_blocks: selection.required_blocks as usize,
-                            overlap_blocks: selection.overlap_blocks,
-                        };
-                        if let Err(e) = hit_rate_publisher.publish(&event).await {
-                            tracing::warn!("Failed to publish KV hit rate event: {:?}", e);
-                        }
-
                         let response = SchedulingResponse {
                             best_worker: selection.worker,
                             overlap_blocks: selection.overlap_blocks,

lib/llm/src/kv_router/sequence.rs

@@ -37,10 +37,10 @@ use std::time::Duration;
 use tokio::time::Instant;
 use uuid::Uuid;
 
+use super::metrics::WORKER_LOAD_METRICS;
 use super::protocols::{
     ActiveLoad, ActiveSequenceEvent, ActiveSequenceEventData, WorkerWithDpRank,
 };
-use crate::discovery::{WORKER_ACTIVE_DECODE_BLOCKS_GAUGE, WORKER_ACTIVE_PREFILL_TOKENS_GAUGE};
 use crate::kv_router::{ACTIVE_SEQUENCES_SUBJECT, KV_METRICS_SUBJECT};
 use crate::local_model::runtime_config::ModelRuntimeConfig;
 use dynamo_runtime::CancellationToken;
@@ -1043,22 +1043,13 @@ impl ActiveSequencesMultiWorker {
         };
 
         // Update Prometheus gauges directly (router's own bookkeeping)
-        let worker_id_str = worker.worker_id.to_string();
-        let dp_rank_str = worker.dp_rank.to_string();
-        WORKER_ACTIVE_DECODE_BLOCKS_GAUGE
-            .with_label_values(&[
-                worker_id_str.as_str(),
-                dp_rank_str.as_str(),
-                self.worker_type,
-            ])
-            .set(active_blocks as i64);
-        WORKER_ACTIVE_PREFILL_TOKENS_GAUGE
-            .with_label_values(&[
-                worker_id_str.as_str(),
-                dp_rank_str.as_str(),
-                self.worker_type,
-            ])
-            .set(active_tokens as i64);
+        WORKER_LOAD_METRICS.observe(
+            worker.worker_id,
+            worker.dp_rank,
+            self.worker_type,
+            active_blocks,
+            active_tokens,
+        );
 
         // Also publish ActiveLoad to NATS for other subscribers (if NATS is available)
         let active_load = ActiveLoad {

lib/llm/src/protocols/common/timing.rs

@@ -6,20 +6,20 @@
 //! This module provides [`RequestTracker`] for tracking timing and routing information
 //! that can be returned to clients via the `nvext` response field.
 
-use std::time::{Instant, SystemTime, UNIX_EPOCH};
-use std::{
-    sync::{
-        Arc, OnceLock,
-        atomic::{AtomicU32, AtomicU64, Ordering},
-    },
-    time::Duration,
-};
+use std::sync::Arc;
+use std::sync::OnceLock;
+use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
+use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
 
 use parking_lot::Mutex;
 use serde::{Deserialize, Serialize};
 use tokio::sync::{OwnedSemaphorePermit, Semaphore};
 use utoipa::ToSchema;
 
+use crate::http::service::metrics::{
+    WORKER_LAST_INPUT_SEQUENCE_TOKENS_GAUGE, WORKER_LAST_INTER_TOKEN_LATENCY_GAUGE,
+    WORKER_LAST_TIME_TO_FIRST_TOKEN_GAUGE,
+};
 use crate::protocols::openai::nvext::WorkerIdInfo;
 
 /// Sentinel value indicating no worker ID has been set.
@@ -62,18 +62,27 @@ impl std::fmt::Display for RequestPhase {
 /// Captures information throughout the request lifecycle:
 /// - `request_received`: When the request was received
 /// - `prefill_start_time`: When prefill started (for disaggregated serving)
-/// - `first_token_time`: When the first token was generated (set once via OnceLock)
-/// - `request_finish_time`: When the request finished (set once via OnceLock)
+/// - `first_token_time`: When the first token was generated
+/// - `request_finish_time`: When the last token was generated (updated incrementally)
 /// - KV cache hit rate information
 /// - Worker IDs and types for per-worker Prometheus metrics
 ///
-/// The `OnceLock` fields ensure that values are set exactly once,
-/// which is important for disaggregated serving where the "first token"
-/// might appear multiple times.
+/// ## Concurrency primitives
+///
+/// **`OnceLock` (first-write-wins):** Used for values that must capture the earliest
+/// observation and ignore later writes. In disaggregated serving, both prefill and decode
+/// phases may call `record_first_token`; `OnceLock` ensures the prefill phase's TTFT is
+/// preserved. Also used for one-shot metadata: `prefill_start_time`, KV hit info,
+/// ISL/cached tokens, worker types, and tokenizer latency.
 ///
-/// Worker IDs use `AtomicU64` instead of `OnceLock<u64>` for lower overhead since
-/// the tracker is created for every request. The sentinel value `NO_WORKER_ID` (0)
-/// indicates no worker has been recorded yet.
+/// **`Mutex` (last-write-wins):** Used for values where later phases should overwrite
+/// earlier ones. `request_finish_time` is updated incrementally at each output block
+/// boundary so that `avg_itl_ms()` stays current during streaming, and the decode
+/// phase's final finish naturally overwrites the prefill phase's earlier finish.
+/// `phase` also uses a Mutex since it transitions across phases.
+///
+/// **`AtomicU64`/`AtomicU32`:** Used for frequently updated counters (`osl_tokens`)
+/// and worker IDs/ranks where `OnceLock`'s heap overhead is unnecessary.
 #[derive(Debug)]
 pub struct RequestTracker {
     /// When the request was received (monotonic clock for duration calculations)
@@ -85,11 +94,14 @@ pub struct RequestTracker {
     /// When prefill started (for disaggregated serving) - set once via OnceLock
     prefill_start_time: OnceLock<Instant>,
 
-    /// When the first token was generated - set once via OnceLock
+    /// When the first token was generated (set once via OnceLock).
+    /// In disaggregated serving, the prefill phase records this first and the
+    /// decode phase's attempt is silently ignored, preserving the real TTFT.
     first_token_time: OnceLock<Instant>,
 
-    /// When the request finished - set once via OnceLock
-    request_finish_time: OnceLock<Instant>,
+    /// When the request finished. Mutex allows the last router phase to
+    /// record the final finish time.
+    request_finish_time: Mutex<Option<Instant>>,
 
     /// KV cache overlap blocks (prefix cache hits) - set once via OnceLock
     kv_overlap_blocks: OnceLock<u32>,
@@ -97,6 +109,15 @@ pub struct RequestTracker {
     /// Input sequence length in blocks (for hit rate calculation) - set once via OnceLock
     isl_blocks: OnceLock<usize>,
 
+    /// Input sequence length in tokens - set once via OnceLock
+    isl_tokens: OnceLock<usize>,
+
+    /// Number of cached tokens (overlap_blocks * block_size) - set once via OnceLock
+    cached_tokens: OnceLock<usize>,
+
+    /// Output sequence length in tokens - updated atomically as tokens stream back
+    osl_tokens: AtomicU64,
+
     /// Prefill worker ID (for disaggregated serving).
     /// Uses atomic with compare-exchange for set-once semantics.
     /// Value of 0 (NO_WORKER_ID) means not yet set.
@@ -128,7 +149,7 @@ pub struct RequestTracker {
     /// Semaphore for coordinating phase transitions.
     /// Acquiring a permit blocks subsequent set_phase calls until the permit is dropped.
     /// This prevents race conditions in the bootstrap optimization path where prefill
-    /// runs in background and needs to complete record_worker before phase changes.
+    /// runs in background and needs to complete record_worker_full before phase changes.
     phase_semaphore: Arc<Semaphore>,
 
     /// How long it took to tokenize the input
@@ -149,9 +170,12 @@ impl RequestTracker {
             request_received_epoch_ms: epoch_ms,
             prefill_start_time: OnceLock::new(),
             first_token_time: OnceLock::new(),
-            request_finish_time: OnceLock::new(),
+            request_finish_time: Mutex::new(None),
             kv_overlap_blocks: OnceLock::new(),
             isl_blocks: OnceLock::new(),
+            isl_tokens: OnceLock::new(),
+            cached_tokens: OnceLock::new(),
+            osl_tokens: AtomicU64::new(0),
             prefill_worker_id: AtomicU64::new(NO_WORKER_ID),
             prefill_dp_rank: AtomicU32::new(NO_DP_RANK),
             decode_worker_id: AtomicU64::new(NO_WORKER_ID),
@@ -169,12 +193,12 @@ impl RequestTracker {
         self.prefill_start_time.set(Instant::now()).is_ok()
     }
 
-    pub fn record_first_token(&self) -> bool {
-        self.first_token_time.set(Instant::now()).is_ok()
+    pub fn record_first_token(&self) {
+        let _ = self.first_token_time.set(Instant::now());
     }
 
-    pub fn record_finish(&self) -> bool {
-        self.request_finish_time.set(Instant::now()).is_ok()
+    pub fn record_finish(&self) {
+        *self.request_finish_time.lock() = Some(Instant::now());
     }
 
     /// Record KV cache hit information. Returns true if this was the first call.
@@ -184,6 +208,29 @@ impl RequestTracker {
         overlap_set && isl_set
     }
 
+    /// Record input sequence length in tokens and cached token count.
+    pub fn record_isl(&self, isl_tokens: usize, cached_tokens: usize) {
+        let _ = self.isl_tokens.set(isl_tokens);
+        let _ = self.cached_tokens.set(cached_tokens);
+    }
+
+    pub fn isl_tokens(&self) -> Option<usize> {
+        self.isl_tokens.get().copied()
+    }
+
+    pub fn cached_tokens(&self) -> Option<usize> {
+        self.cached_tokens.get().copied()
+    }
+
+    /// Record current output sequence length in tokens. Updated at each output block boundary.
+    pub fn record_osl(&self, osl: usize) {
+        self.osl_tokens.store(osl as u64, Ordering::Relaxed);
+    }
+
+    pub fn osl_tokens(&self) -> u64 {
+        self.osl_tokens.load(Ordering::Relaxed)
+    }
+
     /// Time from request received to prefill start (queue/wait time) in milliseconds.
     pub fn prefill_wait_time_ms(&self) -> Option<f64> {
         self.prefill_start_time
@@ -199,15 +246,32 @@ impl RequestTracker {
     }
 
     pub fn ttft_ms(&self) -> Option<f64> {
-        self.first_token_time
-            .get()
-            .map(|t| t.duration_since(self.request_received).as_secs_f64() * 1000.0)
+        let first_token = self.first_token_time.get()?;
+        Some(
+            first_token
+                .duration_since(self.request_received)
+                .as_secs_f64()
+                * 1000.0,
+        )
     }
 
     pub fn total_time_ms(&self) -> Option<f64> {
-        self.request_finish_time
-            .get()
-            .map(|t| t.duration_since(self.request_received).as_secs_f64() * 1000.0)
+        let finish = (*self.request_finish_time.lock())?;
+        Some(finish.duration_since(self.request_received).as_secs_f64() * 1000.0)
+    }
+
+    /// Average inter-token latency in milliseconds.
+    /// Computed as (finish_time - first_token_time) / (osl - 1).
+    /// Returns None if fewer than 2 output tokens or times not recorded.
+    pub fn avg_itl_ms(&self) -> Option<f64> {
+        let first_token = *self.first_token_time.get()?;
+        let finish = (*self.request_finish_time.lock())?;
+        let osl = self.osl_tokens.load(Ordering::Relaxed);
+        if osl < 2 {
+            return None;
+        }
+        let decode_duration = finish.duration_since(first_token).as_secs_f64() * 1000.0;
+        Some(decode_duration / (osl - 1) as f64)
     }
 
     pub fn request_received_epoch_ms(&self) -> u64 {
@@ -224,95 +288,11 @@ impl RequestTracker {
         Some(overlap as f64 / isl as f64)
     }
 
-    /// Record the prefill worker ID. Returns true if this was the first call.
-    pub fn record_prefill_worker(&self, id: u64) -> bool {
-        self.prefill_worker_id
-            .compare_exchange(NO_WORKER_ID, id, Ordering::SeqCst, Ordering::SeqCst)
-            .is_ok()
-    }
-
-    /// Record the prefill worker ID and DP rank. Returns true if worker_id was recorded for the first time.
-    /// Only sets the dp_rank if the worker_id is newly set to avoid mismatched worker_id/dp_rank pairs.
-    pub fn record_prefill_worker_with_rank(&self, id: u64, dp_rank: u32) -> bool {
-        let is_new = self
-            .prefill_worker_id
-            .compare_exchange(NO_WORKER_ID, id, Ordering::SeqCst, Ordering::SeqCst)
-            .is_ok();
-        if is_new {
-            self.prefill_dp_rank.store(dp_rank, Ordering::SeqCst);
-        }
-        is_new
-    }
-
-    /// Record the prefill worker ID, DP rank, and worker type.
-    /// The worker_type is stored to avoid MDC lookup when updating Prometheus metrics.
-    /// Returns true if worker_id was recorded for the first time.
-    pub fn record_prefill_worker_full(
-        &self,
-        id: u64,
-        dp_rank: u32,
-        worker_type: &'static str,
-    ) -> bool {
-        let is_new = self
-            .prefill_worker_id
-            .compare_exchange(NO_WORKER_ID, id, Ordering::SeqCst, Ordering::SeqCst)
-            .is_ok();
-        if is_new {
-            self.prefill_dp_rank.store(dp_rank, Ordering::SeqCst);
-            let _ = self.prefill_worker_type.set(worker_type);
-        }
-        is_new
-    }
-
-    /// Record the decode worker ID. Returns true if this was the first call.
-    pub fn record_decode_worker(&self, id: u64) -> bool {
-        self.decode_worker_id
-            .compare_exchange(NO_WORKER_ID, id, Ordering::SeqCst, Ordering::SeqCst)
-            .is_ok()
-    }
-
-    /// Record the decode worker ID and DP rank. Returns true if worker_id was recorded for the first time.
-    /// Only sets the dp_rank if the worker_id is newly set to avoid mismatched worker_id/dp_rank pairs.
-    pub fn record_decode_worker_with_rank(&self, id: u64, dp_rank: u32) -> bool {
-        let is_new = self
-            .decode_worker_id
-            .compare_exchange(NO_WORKER_ID, id, Ordering::SeqCst, Ordering::SeqCst)
-            .is_ok();
-        if is_new {
-            self.decode_dp_rank.store(dp_rank, Ordering::SeqCst);
-        }
-        is_new
-    }
-
-    /// Record the decode worker ID, DP rank, and worker type.
-    /// The worker_type is stored to avoid MDC lookup when updating Prometheus metrics.
-    /// Returns true if worker_id was recorded for the first time.
-    pub fn record_decode_worker_full(
-        &self,
-        id: u64,
-        dp_rank: u32,
-        worker_type: &'static str,
-    ) -> bool {
-        let is_new = self
-            .decode_worker_id
-            .compare_exchange(NO_WORKER_ID, id, Ordering::SeqCst, Ordering::SeqCst)
-            .is_ok();
-        if is_new {
-            self.decode_dp_rank.store(dp_rank, Ordering::SeqCst);
-            let _ = self.decode_worker_type.set(worker_type);
-        }
-        is_new
-    }
-
     /// Set the request phase and return a permit that blocks subsequent phase changes.
     ///
     /// The returned permit must be dropped to allow the next `set_phase` call to proceed.
-    /// Under normal operation, callers can simply ignore the returned permit (letting it
-    /// drop immediately). In the bootstrap optimization path, the permit is held and
-    /// passed to the spawned prefill task, which drops it after `record_worker` completes.
-    ///
-    /// This prevents the race condition where the phase changes to Decode before the
-    /// background prefill task has recorded its worker ID.
+    /// In the bootstrap optimization path, the permit is held and passed to the spawned
+    /// prefill task, ensuring routing completes before the phase changes.
     pub async fn set_phase(&self, phase: RequestPhase) -> OwnedSemaphorePermit {
         let permit = self
             .phase_semaphore
@@ -329,68 +309,31 @@ impl RequestTracker {
         *self.phase.lock()
     }
 
-    /// Record worker ID based on the current phase.
-    ///
-    /// - Prefill phase: records as prefill_worker_id
-    /// - Decode phase: records as decode_worker_id
-    /// - Aggregated phase: records as both prefill and decode worker
-    pub fn record_worker(&self, instance_id: u64) {
-        match self.phase() {
-            RequestPhase::Prefill => {
-                self.record_prefill_worker(instance_id);
-            }
-            RequestPhase::Decode => {
-                self.record_decode_worker(instance_id);
-            }
-            RequestPhase::Aggregated => {
-                self.record_prefill_worker(instance_id);
-                self.record_decode_worker(instance_id);
-            }
-        }
-    }
-
-    /// Record worker ID and DP rank based on the current phase.
-    ///
-    /// - Prefill phase: records as prefill_worker_id/prefill_dp_rank
-    /// - Decode phase: records as decode_worker_id/decode_dp_rank
-    /// - Aggregated phase: records as both prefill and decode worker/rank
-    pub fn record_worker_with_rank(&self, instance_id: u64, dp_rank: u32) {
-        match self.phase() {
-            RequestPhase::Prefill => {
-                self.record_prefill_worker_with_rank(instance_id, dp_rank);
-            }
-            RequestPhase::Decode => {
-                self.record_decode_worker_with_rank(instance_id, dp_rank);
-            }
-            RequestPhase::Aggregated => {
-                self.record_prefill_worker_with_rank(instance_id, dp_rank);
-                self.record_decode_worker_with_rank(instance_id, dp_rank);
-            }
-        }
-    }
-
     /// Record worker ID, DP rank, and worker type based on the current phase.
     ///
-    /// This is the preferred method when worker_type is known (from MDC or router config),
-    /// as it stores the worker_type for later use in Prometheus metric updates without
-    /// requiring an expensive MDC lookup.
-    ///
-    /// - Prefill phase: records as prefill worker with given worker_type
-    /// - Decode phase: records as decode worker with given worker_type
-    /// - Aggregated phase: records as both prefill and decode worker with the same worker_type
+    /// Each slot is written exactly once by `KvPushRouter::generate()`:
+    /// - Prefill phase: stores as prefill worker
+    /// - Decode phase: stores as decode worker
+    /// - Aggregated phase: stores as both prefill and decode worker
     pub fn record_worker_full(&self, instance_id: u64, dp_rank: u32, worker_type: &'static str) {
         match self.phase() {
             RequestPhase::Prefill => {
-                self.record_prefill_worker_full(instance_id, dp_rank, worker_type);
+                self.prefill_worker_id.store(instance_id, Ordering::Relaxed);
+                self.prefill_dp_rank.store(dp_rank, Ordering::Relaxed);
+                let _ = self.prefill_worker_type.set(worker_type);
             }
             RequestPhase::Decode => {
-                self.record_decode_worker_full(instance_id, dp_rank, worker_type);
+                self.decode_worker_id.store(instance_id, Ordering::Relaxed);
+                self.decode_dp_rank.store(dp_rank, Ordering::Relaxed);
+                let _ = self.decode_worker_type.set(worker_type);
             }
             RequestPhase::Aggregated => {
-                // In aggregated mode, both prefill and decode happen on the same worker,
-                // so we record the same worker_type for both
-                self.record_prefill_worker_full(instance_id, dp_rank, worker_type);
-                self.record_decode_worker_full(instance_id, dp_rank, worker_type);
+                self.prefill_worker_id.store(instance_id, Ordering::Relaxed);
+                self.prefill_dp_rank.store(dp_rank, Ordering::Relaxed);
+                let _ = self.prefill_worker_type.set(worker_type);
+                self.decode_worker_id.store(instance_id, Ordering::Relaxed);
+                self.decode_dp_rank.store(dp_rank, Ordering::Relaxed);
+                let _ = self.decode_worker_type.set(worker_type);
             }
         }
     }
@@ -454,6 +397,51 @@ impl RequestTracker {
         self.decode_worker_type.get().copied()
     }
 
+    /// Write TTFT and ISL to per-worker last gauges using prefill worker labels.
+    /// Called from the Python binding path on first token.
+    pub fn observe_first_token_gauges(&self) {
+        let Some(worker_id) = self.prefill_worker_id() else {
+            return;
+        };
+        let worker_id_str = worker_id.to_string();
+        let dp_rank_str = self
+            .prefill_dp_rank()
+            .map_or("0".to_string(), |r| r.to_string());
+        let worker_type = self.prefill_worker_type().unwrap_or(WORKER_TYPE_PREFILL);
+        let labels = &[worker_id_str.as_str(), dp_rank_str.as_str(), worker_type];
+
+        if let Some(ttft) = self.ttft_ms() {
+            WORKER_LAST_TIME_TO_FIRST_TOKEN_GAUGE
+                .with_label_values(labels)
+                .set(ttft / 1000.0);
+        }
+        if let Some(isl) = self.isl_tokens() {
+            WORKER_LAST_INPUT_SEQUENCE_TOKENS_GAUGE
+                .with_label_values(labels)
+                .set(isl as i64);
+        }
+    }
+
+    /// Write avg ITL to per-worker last gauge using decode worker labels.
+    /// Called at each output block boundary and from the Python binding path.
+    pub fn observe_finish_gauges(&self) {
+        let Some(worker_id) = self.decode_worker_id() else {
+            return;
+        };
+        let worker_id_str = worker_id.to_string();
+        let dp_rank_str = self
+            .decode_dp_rank()
+            .map_or("0".to_string(), |r| r.to_string());
+        let worker_type = self.decode_worker_type().unwrap_or(WORKER_TYPE_DECODE);
+        let labels = &[worker_id_str.as_str(), dp_rank_str.as_str(), worker_type];
+
+        if let Some(avg_itl) = self.avg_itl_ms() {
+            WORKER_LAST_INTER_TOKEN_LATENCY_GAUGE
+                .with_label_values(labels)
+                .set(avg_itl / 1000.0);
+        }
+    }
+
     pub fn get_timing_info(&self) -> TimingInfo {
         TimingInfo {
             request_received_ms: self.request_received_epoch_ms,
@@ -501,3 +489,163 @@ pub struct TimingInfo {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub kv_hit_rate: Option<f64>,
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::thread;
+    use std::time::Duration;
+
+    #[test]
+    fn test_record_isl_osl() {
+        let tracker = RequestTracker::new();
+
+        tracker.record_isl(512, 256);
+        assert_eq!(tracker.isl_tokens(), Some(512));
+        assert_eq!(tracker.cached_tokens(), Some(256));
+
+        tracker.record_osl(100);
+        assert_eq!(tracker.osl_tokens(), 100);
+    }
+
+    #[test]
+    fn test_ttft_ms() {
+        let tracker = RequestTracker::new();
+        thread::sleep(Duration::from_millis(10));
+        tracker.record_first_token();
+
+        let ttft = tracker.ttft_ms().unwrap();
+        assert!(ttft >= 5.0, "TTFT should be at least 5ms, got {ttft}");
+    }
+
+    #[test]
+    fn test_ttft_ms_none_before_first_token() {
+        let tracker = RequestTracker::new();
+        assert!(tracker.ttft_ms().is_none());
+    }
+
+    #[test]
+    fn test_avg_itl_ms() {
+        let tracker = RequestTracker::new();
+        tracker.record_first_token();
+        thread::sleep(Duration::from_millis(20));
+        tracker.record_osl(11); // 11 tokens => 10 inter-token gaps
+        tracker.record_finish();
+
+        let itl = tracker.avg_itl_ms().unwrap();
+        assert!(itl > 0.0, "avg ITL should be positive, got {itl}");
+    }
+
+    #[test]
+    fn test_avg_itl_ms_none_with_single_token() {
+        let tracker = RequestTracker::new();
+        tracker.record_first_token();
+        tracker.record_osl(1);
+        tracker.record_finish();
+
+        assert!(
+            tracker.avg_itl_ms().is_none(),
+            "avg ITL should be None with < 2 output tokens"
+        );
+    }
+
+    #[test]
+    fn test_kv_hit_rate() {
+        let tracker = RequestTracker::new();
+        tracker.record_kv_hit(3, 10);
+
+        let rate = tracker.kv_hit_rate().unwrap();
+        assert!(
+            (rate - 0.3).abs() < f64::EPSILON,
+            "KV hit rate should be 0.3, got {rate}"
+        );
+    }
+
+    #[test]
+    fn test_kv_hit_rate_zero_isl() {
+        let tracker = RequestTracker::new();
+        tracker.record_kv_hit(0, 0);
+        assert!(
+            tracker.kv_hit_rate().is_none(),
+            "KV hit rate should be None when isl_blocks is 0"
+        );
+    }
+
+    #[test]
+    fn test_total_time_ms() {
+        let tracker = RequestTracker::new();
+        thread::sleep(Duration::from_millis(10));
+        tracker.record_finish();
+
+        let total = tracker.total_time_ms().unwrap();
+        assert!(
+            total >= 5.0,
+            "total time should be at least 5ms, got {total}"
+        );
+    }
+
+    #[test]
+    fn test_observe_first_token_gauges_no_panic_without_worker() {
+        let tracker = RequestTracker::new();
+        tracker.record_first_token();
+        tracker.record_isl(100, 50);
+        // No worker recorded — should return early without panic
+        tracker.observe_first_token_gauges();
+    }
+
+    #[test]
+    fn test_observe_finish_gauges_no_panic_without_worker() {
+        let tracker = RequestTracker::new();
+        tracker.record_first_token();
+        tracker.record_osl(10);
+        tracker.record_finish();
+        // No worker recorded — should return early without panic
+        tracker.observe_finish_gauges();
+    }
+
+    #[test]
+    fn test_observe_first_token_gauges_with_worker() {
+        let tracker = RequestTracker::new();
+        tracker.record_worker_full(42, 0, WORKER_TYPE_PREFILL);
+        thread::sleep(Duration::from_millis(5));
+        tracker.record_first_token();
+        tracker.record_isl(256, 128);
+
+        tracker.observe_first_token_gauges();
+
+        let labels = &["42", "0", WORKER_TYPE_PREFILL];
+        let ttft_val = WORKER_LAST_TIME_TO_FIRST_TOKEN_GAUGE
+            .with_label_values(labels)
+            .get();
+        assert!(
+            ttft_val > 0.0,
+            "TTFT gauge should be positive after observe, got {ttft_val}"
+        );
+
+        let isl_val = WORKER_LAST_INPUT_SEQUENCE_TOKENS_GAUGE
+            .with_label_values(labels)
+            .get();
+        assert_eq!(isl_val, 256, "ISL gauge should be 256, got {isl_val}");
+    }
+
+    #[test]
+    fn test_observe_finish_gauges_with_worker() {
+        let tracker = RequestTracker::new();
+        tracker.record_worker_full(99, 1, WORKER_TYPE_DECODE);
+        tracker.record_first_token();
+        thread::sleep(Duration::from_millis(10));
+        tracker.record_osl(5);
+        tracker.record_finish();
+
+        tracker.observe_finish_gauges();
+
+        let labels = &["99", "1", WORKER_TYPE_DECODE];
+        let itl_val = WORKER_LAST_INTER_TOKEN_LATENCY_GAUGE
+            .with_label_values(labels)
+            .get();
+        assert!(
+            itl_val > 0.0,
+            "ITL gauge should be positive after observe, got {itl_val}"
+        );
+    }
+}

lib/llm/src/protocols/openai/chat_completions/delta.rs

@@ -429,11 +429,6 @@ impl crate::protocols::openai::DeltaGeneratorExt<NvCreateChatCompletionStreamRes
             delta.stop_reason,
         );
 
-        // Record first token time (only succeeds on first call due to OnceLock)
-        if let Some(ref tracker) = self.tracker {
-            tracker.record_first_token();
-        }
-
         // Get worker_id info from tracker (set by KvPushRouter based on phase)
         let worker_id_info = self.tracker.as_ref().and_then(|t| t.get_worker_info());
 

lib/llm/src/protocols/openai/completions/delta.rs

@@ -323,11 +323,6 @@ impl crate::protocols::openai::DeltaGeneratorExt<NvCreateCompletionResponse> for
         let index = delta.index.unwrap_or(0);
         let mut response = self.create_choice(index, delta.text.clone(), finish_reason, logprobs);
 
-        // Record first token time (only succeeds on first call due to OnceLock)
-        if let Some(ref tracker) = self.tracker {
-            tracker.record_first_token();
-        }
-
         // Get worker_id info from tracker (set by KvPushRouter based on phase)
         let worker_id_info = self.tracker.as_ref().and_then(|t| t.get_worker_info());
 

lib/runtime/src/metrics/prometheus_names.rs

@@ -68,6 +68,9 @@ pub mod name_prefix {
 
     /// Prefix for frontend service metrics
     pub const FRONTEND: &str = "dynamo_frontend";
+
+    /// Prefix for routing overhead metrics (raw Prometheus, not component-scoped)
+    pub const ROUTING_OVERHEAD: &str = "dynamo_routing_overhead";
 }
 
 /// Automatically inserted Prometheus label names used across the metrics system
@@ -93,6 +96,9 @@ pub mod labels {
     /// It is used by worker/load-style metrics that need to disambiguate per-worker series.
     pub const DP_RANK: &str = "dp_rank";
 
+    /// Label for worker instance ID (etcd lease ID).
+    pub const WORKER_ID: &str = "worker_id";
+
     /// Label for model name/path (OpenAI API standard, injected by Dynamo)
     /// This is the standard label name injected by all backends in metrics_labels=[("model", ...)].
     /// Ensures compatibility with OpenAI-compatible tooling.
@@ -367,6 +373,27 @@ pub mod kvbm {
     pub const OBJECT_WRITE_FAILURES: &str = "object_write_failures";
 }
 
+/// Routing overhead phase latency histogram names (raw Prometheus, not component-scoped).
+///
+/// These are combined with [`name_prefix::ROUTING_OVERHEAD`] to form full metric names,
+/// e.g. `dynamo_routing_overhead_block_hashing_ms`.
+pub mod routing_overhead {
+    /// Time spent computing block hashes
+    pub const BLOCK_HASHING_MS: &str = "block_hashing_ms";
+
+    /// Time spent in indexer find_matches
+    pub const INDEXER_FIND_MATCHES_MS: &str = "indexer_find_matches_ms";
+
+    /// Time spent computing sequence hashes
+    pub const SEQ_HASHING_MS: &str = "seq_hashing_ms";
+
+    /// Time spent in scheduler worker selection
+    pub const SCHEDULING_MS: &str = "scheduling_ms";
+
+    /// Total routing overhead per request
+    pub const TOTAL_MS: &str = "total_ms";
+}
+
 // KvRouter (including KvInexer) Prometheus metric names
 pub mod kvrouter {
     /// Number of KV cache events applied to the index (including status)